Knowledge graph and convolutional neural network fused unreal information detection method

Abstract

The invention provides an unreal information detection method fusing a knowledge graph and a convolutional neural network, and belongs to the technical field of deep learning and natural language processing. According to the method, a knowledge graph is used for supplementing named entities with background missing, deep local semantic representation is carried out on text information embedded with knowledge, meanwhile, the global semantic relation between the information is mined, finally, features are fused, and dynamic features of message propagation are extracted through a long-short term memory network. An unreal information detection model fusing the knowledge graph and the convolutional neural network is formed, and the accuracy of unreal information detection can be well improved through background knowledge supplement, deep semantic information representation and semantic relationship mining between messages; meanwhile, the model is also suitable for early-stage unreal information detection tasks.

Description

Technical Field

[0001] The present invention relates to a false information detection method integrating a knowledge graph and a convolutional neural network, and belongs to the technical field of deep learning and natural language processing. Background Art

[0002] With the rapid development of the Internet, popular online social platforms such as Weibo and Twitter have become the main channels for people to obtain information in their daily lives. Users of these platforms can publish, comment and forward their own or others' speeches relatively freely, breaking the time and space barriers of traditional social interaction. However, while providing convenience for people, online social platforms have also become a breeding ground for false information. Once spread, these false information will disrupt the normal order of society, cause public anxiety, and even cause serious economic losses. Therefore, timely and accurate identification of false information on the Internet will help reduce negative impacts and is of great significance to government departments in improving the efficiency of public management and maintaining social stability.

[0003] Early false information detection required manual intervention to establish feature engineering, and then classify information through decision trees, random forests and other classifiers to make judgments. The features extracted by this method are too simple, resulting in the inability to improve the accuracy of false information detection, and it requires a lot of manpower. In order to achieve more accurate false information detection, neural networks have begun to be applied to false information detection. For example, knowledge graphs are introduced to supplement entities in information, and the conceptual information of text information and entities is constructed into a structural graph and the graph attention neural network is used to extract the semantic features of the message, and then the false information detection model is trained using the features; or the time characteristics of data propagation are taken into account in the model, and the method of false information detection using recurrent neural networks is used to obtain the hidden features of the content of false information that change over time by modeling the propagation data on the time characteristics; or the message is converted into a tree structure, and a recursive neural network is used to capture the message propagation path, and the false information detection effect is improved by modeling the propagation structure.

[0004] Although these false information detection methods using neural network technology can successfully extract information features, there are still some problems, including the lack of background content in the vector representation of short texts, the relatively one-sided expression of semantic information, and the lack of semantic relationships between information, which can easily cause the lack of key feature content and affect the detection results. Therefore, the accuracy of online false information detection needs to be improved. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention proposes a false information detection method that integrates knowledge graphs and convolutional neural networks, which improves the accuracy of false information detection on the Internet by supplementing background knowledge of text information, extracting local and global semantic features of information in different time periods, and extracting deep feature relationships through attention mechanisms.

[0006] Technical solution of the present invention: In recent years, research in related fields has mainly used neural networks to extract feature representations. Models trained using nonlinear features extracted from neural networks can significantly improve the accuracy of false information detection. The present invention mainly introduces related research work from two aspects: the structural characteristics of false information and the fusion of external knowledge.

[0007] The present invention provides a false information detection method integrating knowledge graph and convolutional neural network, comprising the following steps: Step 1: Collect source post information of events spread in social media, where the information provided by social media mainly includes event set, source post, comment, forwarding, and time information; Step 2: Build a dynamic graph convolutional false information detection model that integrates the knowledge graph to expand the text content, which specifically includes the following steps: Step 2.1, construct a knowledge graph, and introduce the knowledge graph to supplement the information with background knowledge; Step 2.2: Use text preprocessing technology to vectorize the information that incorporates background knowledge, and then convert the vector form of the processed source post information and comment forwarding information into a matrix form and input it into the convolutional neural network to extract semantic features and extract local semantic relationships; Step 2.3: Introduce graph convolutional neural network and extract global semantic relations through cross-attention mechanism; Step 2.4, dynamic feature fusion; Step 2.5, optimize the false information detection classifier; Step three: Test the model performance on a public data set and output the false information detection results.

[0008] As a further illustration of the present invention, in step 2.1, the specific process of using the knowledge graph to supplement the information with background knowledge is as follows: First, entities in the text are extracted through entity recognition. The entities in the open domain include names of people, names of organizations / institutions, geographic locations, time / dates, character values, and monetary values. Next, CN-DBpedia and DBpedia are used to link Chinese and English texts to the knowledge graph to capture the explanatory information of related entities. Finally, the explanatory information and text information of the entities are converted into graph form.

[0009] As a further illustration of the present invention, in step 2.2, the specific operation flow of the local semantic relationship extraction is as follows: with the help of knowledge graph technology, the fuzzy entities in each piece of information are linked to the knowledge graph for information matching, and the content interpretation of the entity is obtained, and then the interpretation content is embedded in the text to supplement the original text semantics; then the text content embedded in the interpretation information is vectorized using the pre-trained model BERT, and the convolutional neural network is used to extract the deep semantic features of the message from the event representation.

[0010] As a further explanation of the present invention, in step 2.4, the specific operation process of the dynamic feature fusion is as follows: the overall time of message propagation is evenly divided into m time periods, the corresponding semantic features are extracted in each time period, and then the features in each time period on the entire timeline are spliced, and finally the time series features are input into the model to extract the overall dynamic features of the information.

[0011] As a further illustration of the present invention, a long short-term memory network is used to extract the temporal features. Beneficial Effects

[0012] The present invention proposes a false information detection method that integrates knowledge graph and convolutional neural network. The method uses knowledge graph to supplement the named entities with missing background, and performs deep local semantic representation of text information embedded with knowledge, while mining the global semantic relationship between information, and finally fuses the features and uses long short-term memory network to extract the dynamic characteristics of message propagation, forming a false information detection model that integrates knowledge graph and convolutional neural network. Through the supplement of background knowledge, the representation of deep semantic information and the mining of semantic relationship between messages, it can be judged whether the source post information of a certain event is false information, which greatly improves the accuracy of false information detection; at the same time, this model is also suitable for early false information detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a false information detection method integrating knowledge graph and convolutional neural network of the present invention; Figure 2 for Figure 1 The technical route for false information detection; Figure 3 for Figure 2 The process of introducing external knowledge into the system; Figure 4 for Figure 2 The text vector representation in ; Figure 5 for Figure 2 The LSTM workflow in . DETAILED DESCRIPTION

[0014] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It will be appreciated that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the inventions. It should also be noted that, for ease of description, only portions related to the relevant inventions are shown in the accompanying drawings. It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0015] The present invention provides a false information detection method integrating knowledge graph and convolutional neural network, comprising the following steps: Step 1: Collect source post information of events spread in social media. The information provided by social media mainly includes event set, source post, comment, forwarding, time and other information. The present invention mainly detects false information about events spread in social media, and determines whether the source post information of a certain event is false information rather than whether a certain discussion information in a comment or forwarding information is false information. The information provided by social media mainly includes event sets, source posts, comments, forwarding, time and other information. Let E = {E1, E2, E3…E n} represents the set of events of false information dissemination, where n represents the number of events in the dataset. Each event contains the source post, comments, and release time information, so we represent the text content as Among them, r0 represents the source post content, represents the content of the ith comment of the ith event, t i Indicates the time when the i-th information is released.

[0016] The task of false information detection can be regarded as a binary classification problem, which is actually training a model f:E i →y, each event E i Classify into the corresponding label y, where y∈(0,1), 0 indicates that the event label is not false information, and 1 indicates that the event label is false information.

[0017] Step 2: Build a dynamic graph convolutional false information detection model that integrates the knowledge graph to expand the text content.

[0018] Since comment posts are related discussions based on the content of the source post, it is necessary to combine the source post and comment information to mine the semantic expression of the information. At the same time, the structure of the message will change over time during the spread of false information. The messages are divided based on time and the structural changes are dynamically examined to fully mine the semantic relationship between messages and obtain richer semantic features. The key technologies involved in this study include knowledge graph construction, local semantic relationship extraction, global semantic relationship extraction, and dynamic feature fusion. Specific technical routes are as follows: Figure 2 shown.

[0019] Step 2.1, build a knowledge graph.

[0020] By expanding the entity knowledge of short text information through external knowledge bases, on the one hand, it can enhance the semantic understanding of the text, and on the other hand, it can alleviate the problem of data sparsity. Entity knowledge can expand the amount of information in the text and mine potential semantic representations, providing information supplement for the judgment of false information. The construction of vector graphs can mine the structural relationship between different entities, making the expression of information dissemination digital and easy for computer processing. The specific process of using knowledge graphs to supplement information is as follows: Figure 3 shown.

[0021] First, entities in the text are extracted through entity recognition. Open domain entities include names of people, organization / institution names, geographic locations, time / dates, character values, and monetary values. Next, CN-DBpedia and DBpedia are used to link Chinese and English texts to the knowledge graph to capture the explanatory information of related entities. Finally, the explanatory information and text information of the entities are converted into vector graphs. For example, for a short text message "The more half-moons on the nails, the healthier the body, and the fewer half-moons, the worse the body?", entity recognition is used to generate an entity set T = {nail, half-moon, health}. Then the knowledge graph is used to conceptualize and extract the entity set c 指甲 ={keratin, connective tissue}, c 月牙 ={lunula, lunula marks, keratinocytes}, c 健康 = {physical condition, good spirit}. Finally, the concepts of each entity are combined into a concept set, that is, C = {c1, c2, ..., c n}.

[0022] The representation form of the vector graph is G = (V, E), where V represents a node set and E represents an edge set. The node set in the present invention mainly includes source post information, comment forwarding information, concept set, and entity set, and the edge is represented by the relationship between each node. The edge relationship between source post-comment forwarding, concept-concept, concept-entity, and entity-entity involved in the present invention is calculated using point mutual information (PMI). The relationship between information-entity edges is set according to the term frequency-inverse document frequency (TF-IDF) of the entity in the post, where the term frequency is the number of times the entity appears in the post, and the inverse document frequency is the inverse score of the logarithm of the number of posts that can be connected to the entity.

[0023] The PMI value reflects the correlation between nodes, and a positive value means a high semantic correlation. Therefore, for the vector graph adjacency matrix A i,j Only edges with positive PMI and TF-IDF scores are retained.

[0024] Step 2.2: local semantic relationship extraction.

[0025] Since network information is mainly presented in a concise manner, the present invention adds a knowledge graph to supplement the content of the disseminated information. At the same time, considering that there is a certain connection between the source post information and the comment information, only performing a separate semantic compilation on the source post information or the comment information will lose the connection between the information, resulting in the computer's insufficient understanding of the text semantics. To this end, the present invention designs a background knowledge-enhanced semantic representation to perform feature representation of text information, and mines the connection between words and messages to improve the computer's understanding of information. This section mainly uses text preprocessing technology to perform vector representation on information that incorporates background knowledge, and then converts the vector form of the processed source post information and comment forwarding information into a matrix form and inputs it into the convolutional neural network to extract semantic features. The specific operation process is as follows: Figure 4 shown.

[0026] With the help of knowledge graph technology, the fuzzy entities in each piece of information are linked to the knowledge graph for information matching, the content interpretation of the entity is obtained, and then the interpretation content is embedded in the text to supplement the semantics of the original text. The present invention takes a piece of false information as an example: "With the gradual enhancement of environmental awareness, more and more people have begun to pay attention to new energy vehicles. However, some people believe that new energy vehicles are more polluting than fuel vehicles." First, the entities "new energy vehicles" and "fuel vehicles" of this sentence are extracted, and then the two entities are explained respectively through the knowledge graph: "New energy vehicles refer to vehicles that use unconventional vehicle fuels as power sources, and do not produce any exhaust gas during driving"; "Fuel vehicles refer to vehicles that use oil as fuel to drive vehicles, and their exhaust gas contains pollutants such as volatile organic compounds, nitrogen oxides and sulfur dioxide." Then the entity explanation information is embedded in the original text, that is, "With the gradual enhancement of environmental awareness, more and more people have begun to pay attention to new energy vehicles. However, some people believe that new energy vehicles (new energy vehicles refer to vehicles that use unconventional vehicle fuels as power sources, and do not produce any exhaust gas during driving) are more polluting than fuel vehicles (fuel vehicles refer to vehicles that use oil as fuel to drive vehicles, and their exhaust gas contains pollutants such as volatile organic compounds, nitrogen oxides and sulfur dioxide)."

[0027] Then, the text content embedded with the explanation information is represented by vectors using the pre-trained model BERT. In order to make full use of the information provided by social networks, the present invention not only supplements the source post information with explanations of relevant content, but also uses the forwarding information under each source post to embed explanations for the forwarding information. Since simple concatenation of vectors cannot capture the deep semantic information of the text, the present invention uses a convolutional neural network to extract the deep semantic features of the message from the event representation. This method not only captures the connection between the source information and the comments, but also mines the local relationship between words.

[0028] Specifically, let the source text incorporating background knowledge be represented by r'0, and the comment information incorporating background knowledge be represented by Represented by, after processing by the BERT model, we get the feature vector R of the source text and the feature vector C of the comment information respectively. i Then, concatenate each comment vector row by row to form a matrix T i =[C1; C2…C n ], and then perform convolution operations on the source post vector and comment matrix to extract features: Where W∈R h×d is the parameter matrix of the convolution kernel h×d, and σ is the sigmoid activation function.

[0029] Get the source post feature map And comment information feature map Finally, the maximum pooling is used to reduce the dimension of the features extracted by convolution and extract the more important parameter information in the feature matrix. Assume that O represents the source post information feature representation, M i Representing the comment feature representation, considering that there is a connection between the source post information and the comment information, different information provides different key factors. In order to effectively capture the semantic relationship between information, the present invention chooses to use the cross-attention mechanism to mine the semantic connection between information. M i =atten OM *M V (6) Among them, O Q ,M K ,M V They represent the query conversion of source post features, the key conversion of comment features, and the value conversion of comment features respectively. d is the vector dimension. The final text vector of an event is represented by E i =[O; M1; M2…Mn ].

[0030] Step 2.3, global semantic relationship extraction.

[0031] The spatial structure of false information will change over time during its propagation, so the present invention introduces a graph convolutional network to extract the global semantic feature relationship of information. Considering that background knowledge also changes dynamically as the message spreads, the present invention adds knowledge graph information when extracting the global semantic feature relationship. From step 2.1, the vector graph G and adjacency matrix A of the propagated information can be obtained, and the propagation semantic features are obtained through a two-layer graph convolutional neural network, and the formula is as follows: Among them, H 0 =X, where X represents the matrix of all node features in the propagation graph. 0 ,W 1 ,b 0 ,b 1 represents the trainable parameter matrix. σ(·) represents the nonlinear activation function ReLU, represents the normalized symmetric adjacency matrix, defined as: in, D is the degree matrix, and I is an N-dimensional identity matrix.

[0032] Step 2.4, dynamic feature fusion.

[0033] Taking into account the temporal characteristics of message propagation, the present invention divides the overall time of message propagation into m time periods on average, extracts corresponding semantic features in each time period, then splices the features in each time period on the entire time axis, and finally inputs them into the temporal feature extraction model to extract the overall dynamic features of the information. The splicing formula of local semantic and global semantic features is as follows: X=Concat(E,H) (10) The networks used for temporal feature extraction in neural networks mainly include RNN, LSTM, GRU, etc. Compared with the traditional RNN structure, LSTM (Long Short-Term Memory Network) can effectively solve the long sequence problem by introducing the concepts of memory cells, input gates, output gates and forget gates. Although GRU is simpler than LSTM structure, for large data set detection tasks such as text processing, too few parameters will lead to model overfitting. Therefore, the present invention uses LSTM network to extract temporal features. Its workflow is as follows Figure 5As shown in the figure, i is the input gate: determines whether to write the current input information into the memory cell; f is the forget gate: determines whether to forget the information in the memory cell; o is the output gate: determines whether to use the information in the memory cell as the current output. The control of these gates can effectively capture the important long-term dependencies in the sequence and solve the gradient problem.

[0034] The calculation formulas for each gate are as follows: i m =σ(W xi X m +W hi h m-1 +b i )(11) f m =σ(W xf X m +W hf h m-1 +b f )(12) o m =σ(W xo X m +W ho h m-1 +b o )(13) g m =φ(W xc x m +W hc h m-1 +b c )(14) c m =f m ⊙c m-1 +i m ⊙g m (15) h m =o m ⊙φ(c t )(16) Among them, X m represents the message semantic feature matrix in the mth time period, W and b belong to the parameter matrix, σ(·) is the sigmoid function, φ(·) is the tanh function, c m Represents a memory unit, h m is the hidden state. The hidden states of each time period in the entire propagation process are concatenated to obtain H'={h1,h2,…h l}

[0035] Step 2.5, optimize the false information detection classifier.

[0036] The information representation of the last output hidden state input into the fully connected layer aggregation node is: S = tanh(W*H′+b) (17 Finally, the label of the input information is calculated through the fully connected layer and activation function. Assume that the false information label is 1 and the non-false information label is 0, and use the cross entropy loss function to optimize the false information classifier. That is:

[0037] Step three: Test the model performance on a public data set to obtain false information detection results.

[0038] This paper selects two public data sets, Weibo and Pheme, to conduct model performance test experiments. The data sets come from the domestic Weibo social platform and the foreign Twitter social platform, respectively. The authenticity of the messages has been verified by official accounts. Each event label is false information or not. The data set content mainly includes the original content, comment content, message sending time, etc. The detailed statistical data of the data set are shown in Table 1.

[0039] Table 1 Dataset statistics

[0040] Experimental environment: Windows×11_64; GTX3050; python3.8; pytorch1.2.0

[0041] Experimental details: The present invention uses CN-DBpedia and DBpedia to capture knowledge graph information for the Chinese Weibo dataset and the English Pheme dataset, respectively, and uses the pre-trained model BERT to vectorize the text, where the text length is set to 30, the word vector dimension is set to 768, and the graph convolutional neural network is used to extract the propagation semantic features of the message. The output dimension of the convolutional layer is 128, the number of convolutional layers is set to 2, and the CNN network is selected for the extraction of deep semantic features, where the convolution kernel size is (3, 4, 5). Other hyperparameters such as the learning rate are set to 0.0005, the dropout is set to 0.4, the batch size is 128, and the number of training times is 300. Set the early stopping mechanism, and select Adam as the optimizer to find the optimal network parameters.

[0042] According to the knowledge graph, semantic features, and temporal features involved in the present invention, the following groups of comparison models are selected for performance comparison. (1) BERT: Only the text content of the input information is considered, and the BERT model is used for feature extraction. Then the output feature vector is directly input into the classifier for false information detection; (2) att-RNN: Add attention mechanism to the multimodal fusion process, generate time step features that jointly represent text features and context features through LSTM as the input of attention features, and fuse attention features and visual features into the classifier for false information judgment. (3) CNN: Learn the deep semantic information of text through convolutional neural networks, and explore the potential relationship between each microblog, thereby avoiding structural problems; (4) RvNN: Convert the post into a tree structure, then model a recursive neural network to learn the vector representation of the post, and finally use a classifier to determine whether the post is false information. (5) BiGCN: A bidirectional GCN using a top-down graph convolutional network (TD-GCN) and a bottom-up graph convolutional network (BU-GCN) is used to obtain the propagation and diffusion characteristics of information, and then these two features are used for false information detection. (6) KMGCN: Knowledge-driven multimodal graph convolutional network model for false information detection. It models semantic representation by jointly modeling text information, knowledge concepts, and visual information into a unified fake news detection framework. The text content is constructed into a graph structure, and knowledge concepts are added as node supplementary information. Then, graph convolution operations are performed to extract features for false information detection. (7) DDGCN: A new dual dynamic graph convolutional model for false information detection that combines the dynamic structural features of message propagation and the dynamic features of background knowledge for false information detection.

[0043] In machine learning, the evaluation indicators for the performance of classification models mainly include accuracy (Acc), precision (Pre), recall (Rec) and F1 score. Therefore, the models involved in the present invention all select these four indicators as the basis for comparison.

[0044] Table 2 Experimental results of the dataset

[0045] From the experimental results in Table 2, we can find that the accuracy of experiments on the two public datasets Weibo and Pheme reached 95.7% and 88.4% respectively. All indicators of KG-CNN are better than other baseline models, which shows that incorporating background knowledge and semantic relationships is conducive to improving the accuracy of false information detection.

[0046] BERT, as a model that relies only on text information for false information detection, has the lowest evaluation index among all models. This result shows that in the complex process of false information propagation, it is impossible to accurately identify false information by relying solely on text content. The att-RNN model takes into account the temporal characteristics of message propagation, so it adds an LSTM network as a time series processing unit to integrate temporal features into the false information detection model. Its experimental results are significantly higher than those of the BERT model. This result shows that the temporal characteristics of false information propagation have an impact on the accuracy of false information detection, but its index is lower than that of KG-CNN, indicating that in addition to temporal features, there are other features that affect the accuracy of the model. The index of the CNN model is higher than that of the BERT model, indicating that there is a certain connection between messages and between words. From the experimental results of the two models RvNN and BiGCN that consider the propagation structure, it can be found that the structural features have a greater impact on the accuracy of false information detection than the temporal features. The main reason is that false information is mainly propagated in the form of conversations and therefore relies on the connection between communities. The accuracy of the BiGCN model is higher than that of the RvNN model, indicating that false information propagation is not only horizontal but also vertical, so converting the propagation structure into a graph structure can more accurately express the connection between information. In order to solve the problem of missing background in short texts, the KMGCN model introduces knowledge graphs to enrich the content expression of text information. The results show that the accuracy of the false information detection model with the addition of knowledge graphs has indeed improved. DDGCN not only takes into account the knowledge graph, but also the dynamic characteristics of message propagation, so its model is better than KMGCN. In addition to the features mentioned above, the KG-CNN model also adds semantic features and a temporal memory network. The results show that the semantic features and temporal features of text information after the fusion of the knowledge graph affect the accuracy of the detection model.

[0047] Based on the above analysis, the background knowledge, enhanced semantic features, temporal features, and semantic feature relationships added to this model have a certain impact on the semantic expression of information and the mining of relationships between structures. Compared with the latest DDGCN model, its accuracy has increased by 3.2%, precision by 4.6%, recall by 1.2%, and F1 by 2.9%.

[0048] In order to verify the effect of each module on this model, the present invention designed an ablation experiment to analyze each module separately. KG-CNN / knowledge means removing the knowledge graph; KG-CNN / CNN means removing the semantic enhancement; KG-CNN / propagation means removing the propagation structure; KG-CNN / temporal means removing the temporal fusion module.

[0049] Table 3 Ablation experiment results

[0050] Table 3 presents the results of all ablation experiments. It can be seen from the table that removing any module will reduce the evaluation index of the original model. Therefore, the modules involved in this experiment affect the effect of false information detection.

[0051] The results show that removing structural features has the greatest impact on the accuracy of the model, and the accuracy rate is reduced by 8.5%. It can be seen that the structural features of false information dissemination are an important reference for verifying false information. It shows that there is a connection between each piece of information during the dissemination of false information. Since false information is spread in the form of conversations, a source post information can be commented on by multiple users, and this hidden structural relationship can provide clues for false information detection. In addition to structural features, the present invention also verifies the influence of time features on false information detection. From the results, it can be found that the time features of false information also have an impact on the detection results. This is because as time goes by, false information is also exposed and corrected by informed people while spreading. In the process of spreading false information, netizens in the comment area will raise questions and objections, which will affect the spread of false information. Therefore, the longer the time of dissemination, the easier it is to verify the authenticity of the content. After removing the knowledge graph module, its accuracy rate decreased by 3.1%. The result shows that the introduction of knowledge graphs can supplement the content expression of texts, and complete text expressions can improve the computer's understanding of texts. Although removing the CNN module has the least impact on the original model, its effect is still not as good as the original model. This result also shows that there are connections between posts and between words.

[0052] In summary, the modules involved in this model play a certain role in extracting the propagation structure, dynamic changes, and semantic representation of false information, and significantly improve the efficiency of model detection.

[0053] Through the ablation experiment results, it can be found that the time feature has an impact on the accuracy of false information detection. The spread of false information gradually spreads over time. In the early stages of the spread, the negative impact of false information on society is relatively small. When it spreads widely, the harm of false information to society will become serious. Therefore, detecting false information in the early stages of message propagation can prevent the negative impact of false information. For this reason, the present invention uses early false information detection as another indicator to evaluate the effect of the model.

[0054] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A false information detection method integrating knowledge graph and convolutional neural network, characterized in that: The following steps are included: Step 1: Collect source post information of events spread in social media, where the information provided by social media mainly includes event set, source post, comment, forwarding, and time information; Step 2: Build a dynamic graph convolutional false information detection model that integrates the knowledge graph to expand the text content, which specifically includes the following steps: Step 2.1, construct a knowledge graph, and introduce the knowledge graph to supplement the information with background knowledge; Step 2.2: Use text preprocessing technology to vectorize the information that incorporates background knowledge, and then convert the vector form of the processed source post information and comment forwarding information into a matrix form and input it into the convolutional neural network to extract semantic features and extract local semantic relationships; Step 2.3: Introduce graph convolutional neural network and extract global semantic relations through cross-attention mechanism; Step 2.4, dynamic feature fusion; Step 2.5, optimize the false information detection classifier; Step three: Test the model performance on a public data set and output the false information detection results.

2. The false information detection method according to claim 1, characterized in that: In step 2.1, the specific process of using the knowledge graph to supplement the information with background knowledge is as follows: First, entities in the text are extracted through entity recognition. The entities in the open domain include names of people, names of organizations / institutions, geographic locations, time / dates, character values, and monetary values. Next, we use CN-DBpedia and DBpedia to link Chinese and English texts to the knowledge graph to capture the explanation information of related entities. Finally, the entity’s explanatory information and textual information are converted into a graph.

3. The false information detection method according to claim 2, characterized in that: In step 2.2, the specific operation process of the local semantic relationship extraction is as follows: with the help of knowledge graph technology, the fuzzy entities in each piece of information are linked to the knowledge graph for information matching, the content interpretation of the entity is obtained, and then the interpretation content is embedded in the text to supplement the original text semantics; then the text content embedded in the interpretation information is vectorized using the pre-trained model BERT, and the convolutional neural network is used to extract the deep semantic features of the message from the event representation.

4. The false information detection method according to claim 3, characterized in that: In step 2.4, the specific operation process of the dynamic feature fusion is as follows: divide the overall time of message propagation into m time periods on average, extract the corresponding semantic features in each time period, then splice the features in each time period on the entire timeline, and finally input the time series features into the model to extract the overall dynamic features of the information.

5. The false information detection method according to claim 4, characterized in that: A long short-term memory network is used to extract the temporal features.

Images